AGENT SKILL

agents

Implement and debug NAVSIM rule-based, privileged, learned, TransFuser, and latent TransFuser agents while preserving sensor, trajectory, feature-builder, and checkpoint contracts.

License 许可
Apache-2.0
最近更新
2026-09-03
Stars
★ 351
Forks
27

01 前言 Front Matter

name
agents
license
Apache 2.0
metadata
{ "disco-role": "operating" }
description
Implement and debug NAVSIM rule-based, privileged, learned, TransFuser, and latent TransFuser agents while preserving sensor, trajectory, feature-builder, and checkpoint contracts.
disable-model-invocation
true

02 资源文件 5

03 SKILL.md 正文

NAVSIM agents

Use this route when an agent must be selected, implemented, trained, loaded, or troubleshot. It covers the AbstractAgent interface, rule-based and privileged baselines, ego-status and TransFuser/LTF learned agents, sensor history, local trajectory sampling, feature/target builders, and checkpoint loading.

Route by intent

Required operating rules

  1. Subclass AbstractAgent, expose a stable name(), initialize once per worker, return a deliberate SensorConfig, and return local rear-axle BEV (x, y, heading) poses in a Trajectory.
  2. Treat the trajectory sampling object as part of the model contract. The repository default requests a 4-second horizon sampled every 0.5 seconds; evaluation uses a 4-second, 0.1-second proposal. Do not silently change one to satisfy the other.
  3. A submission receives AgentInput, not an annotated Scene. Keep requires_scene=False for submission agents; the human/privileged baseline is useful for analysis but is prohibited in submission creation.
  4. Choose sensors explicitly. TransFuser uses the current front camera trio and current LiDAR by default; latent TransFuser replaces the LiDAR branch with a learned latent tensor and must not request or build a LiDAR feature.
  5. Learned agents must keep feature-builder keys, target-builder outputs, forward() output (including "trajectory" with shape [B, T, 3]), loss, optimizer, and checkpoint key namespaces mutually compatible.

Safe verification

Run the bundled agent contract validator:

python scripts/validate_agent_contract.py
python scripts/validate_agent_contract.py --help

The validator uses synthetic arrays and sensor configurations only. It does not require a dataset, maps, checkpoints, downloads, submission credentials, or a benchmark. Add --module path/to/your_agent.py to inspect importable agent classes without instantiating or running them; use --strict when missing optional NAVSIM dependencies should fail the check.

Keep training, scoring, submission generation, and dataset access on their respective workflow routes. This sub-skill explains how an agent participates in those workflows but does not launch them by default.

04 同仓库更多 Skills

VectorSpaceLab
VectorSpaceLab/AREX-Skill

passnet-triton-opt

Make a single PassNet Triton kernel fast AND numerically faithful: performance model (when a replacement can win at all), block/grid/warp tuning, autotune policy, launch overhead, and per-op numeric recipes to pass the dtype baseline tolerances. Use when a pass matches and is correct but speedup below expected, or when correctness fails by small numeric margins.

★ 351
VectorSpaceLab
VectorSpaceLab/AREX-Skill

passnet-solve

END-TO-END playbook for solving one PassNet sample: analyze the computation graphs, decide the optimization strategy (what to fuse, what to replace, what to leave alone), drive the iteration loop, and maximize the sample score. This is the ENTRY skill — invoke it first for any PassNet optimization task; it tells you when to use passnet-pattern-fusion, passnet-triton-opt and passnet-feedback.

★ 351
VectorSpaceLab
VectorSpaceLab/AREX-Skill

passnet-skill

PassNet GPU kernel optimization via compiler passes. Design and implement Triton-based optimization passes, create pass files under ./pass_dir/, self-evaluate with pass_evaluator, and iterate to maximize GPU speedup.

★ 351
VectorSpaceLab
VectorSpaceLab/AREX-Skill

passnet-pattern-fusion

Author PassNet pass files: write patterns that actually MATCH the FX graph, pick fusion regions (which ops to absorb into one Triton kernel), and structure multi-pass submissions with the mandatory shared-dispatch architecture. Use when creating or fixing pattern()/replacement_args()/replacement_func() files, when a pass "failed to match", or when deciding how to fuse multiple kernels/ops.

★ 351
VectorSpaceLab
VectorSpaceLab/AREX-Skill

passnet-orchestrate

Round/batch planning and decision support for PassNet graph optimization. Use for multi-sample triage, worker allocation, eval-budget policy, and post-eval keep/revert decisions. For solving one concrete sample end to end, use passnet-solve as the entry skill; consult this skill only when you need broader planning or a specific decision gate.

★ 351
VectorSpaceLab
VectorSpaceLab/AREX-Skill

passnet-feedback

Fast, deterministic feedback for PassNet work: pre-flight pattern/match verification WITHOUT burning a GPU evaluation, per-node bottleneck analysis of a sample's graphs, and eval-log parsing into per-variant status + estimated score + failure classification. Use BEFORE every GPU evaluation (check_pattern), at the START of a sample (analyze_graph), and AFTER every evaluation (parse_eval_log).

★ 351